RE: [PATCH v3 02/15] xen/x86: introduce new sub-hypercall to propagate CPPC data

["Penny, Zheng" <penny.zheng@xxxxxxx>]

[Public]

Hi,

> -----Original Message-----
> From: Jan Beulich <jbeulich@xxxxxxxx>
> Sent: Monday, March 24, 2025 10:28 PM
> To: Penny, Zheng <penny.zheng@xxxxxxx>
> Cc: Huang, Ray <Ray.Huang@xxxxxxx>; Andrew Cooper
> <andrew.cooper3@xxxxxxxxxx>; Roger Pau Monné <roger.pau@xxxxxxxxxx>;
> Anthony PERARD <anthony.perard@xxxxxxxxxx>; Orzel, Michal
> <Michal.Orzel@xxxxxxx>; Julien Grall <julien@xxxxxxx>; Stefano Stabellini
> <sstabellini@xxxxxxxxxx>; xen-devel@xxxxxxxxxxxxxxxxxxxx
> Subject: Re: [PATCH v3 02/15] xen/x86: introduce new sub-hypercall to 
> propagate
> CPPC data
>
> On 06.03.2025 09:39, Penny Zheng wrote:
> > +    pm_info = processor_pminfo[cpuid];
> > +    /* Must already allocated in set_psd_pminfo */
> > +    if ( !pm_info )
> > +    {
> > +        ret = -EINVAL;
> > +        goto out;
> > +    }
> > +    pm_info->cppc_data = *cppc_data;
> > +
> > +    if ( cpufreq_verbose )
> > +        print_CPPC(&pm_info->cppc_data);
> > +
> > +    pm_info->init = XEN_CPPC_INIT;
>
> That is - whichever Dom0 invoked last will have data recorded, and the other
> effectively is discarded? I think a warning (perhaps a one-time one) is 
> minimally
> needed to diagnose the case where one type of data replaces the other.
>

In last v2 discussion, we are discussing that either set_px_pminfo or 
set_cppc_pminfo shall be invoked,
which means either PX data is recorded, or CPPC data is recorded.
Current logic is that, cpufreq cmdline logic will set the 
XEN_PROCESSOR_PM_PX/CPPC
flag to reflect user preference, if user defines the fallback option, like 
"cpufreq=amd-cppc,xen", we will have both
 XEN_PROCESSOR_PM_PX | XEN_PROCESSOR_PM_CPPC set in the beginning.
Later in cpufreq driver register logic, as only one register could be 
registered , if amd-cppc
being registered successfully, it will clear the  XEN_PROCESSOR_PM_PX flag bit.
But if it fails to register, fallback scheme kicks off, we will try the legacy 
P-states, in the mean time,
clearing the XEN_PROCESSOR_PM_CPPC.
We are trying to make XEN_PROCESSOR_PM_PX and XEN_PROCESSOR_PM_CPPC exclusive
values after driver registration, which will ensure us that either 
set_px_pminfo or set_cppc_pminfo
is taken in the runtime.

> With this it also remains unclear to me how fallback to the legacy driver is 
> intended
> to be working. Both taken together are a strong suggestion that important
> information on the model that is being implemented is missing from the 
> description.
>
> > @@ -27,8 +28,6 @@ struct processor_performance {
> >      struct xen_pct_register status_register;
> >      uint32_t state_count;
> >      struct xen_processor_px *states;
> > -
> > -    uint32_t init;
> >  };
> >
> >  struct processor_pminfo {
> > @@ -37,6 +36,9 @@ struct processor_pminfo {
> >      struct xen_psd_package domain_info;
> >      uint32_t shared_type;
> >      struct processor_performance    perf;
> > +    struct xen_processor_cppc cppc_data;
> > +
> > +    uint32_t init;
> >  };
>
> This moving of the "init" field and the mechanical changes coming with it can 
> likely
> be split out to a separate patch? Provided of course the movement is still
> wanted/needed with patch 1 re-worked or dropped.
>
> Jan